Benchmark Almanac

Malicious computer use (Anthropic)

Claude Sonnet 4.6 99.4% leads since Feb 2026 Only in system cards

Computer-use agents receive harmful tasks such as surveillance or fraud in GUI environments. Score is the refusal rate.

In-house Anthropic evaluation. The tasks are not public.

Made by Anthropic Released Nov 2025 Refusing harmful requests Tasks kept private Last in a system card Oct 2026 Last model added Oct 2026
Level
Everyday Most adults, without training
Scored
Graded by a model A language model scores each response, usually against a rubric.
Answers
Agent task Many steps in a terminal, browser, desktop, app, game or robot.
13
Models scored in lab cards
17
System cards report it, from 1 lab

Claude Sonnet 4.6 leads with 99.4%

Best score per model as labs report them in their own cards. Labs pick their own settings, so these compare less cleanly than independent runs.

13 models. Scroll sideways for the rest.

  1. 99.4 Claude Sonnet 4.6
  2. 93.8 Claude Mythos Preview
  3. 93.8 Claude Opus 5
  4. 89.3 Claude Opus 4.7
  5. 88.4 Claude Opus 4.5
  6. 88.3 Claude Opus 4.6
  7. 85.7 Claude Fable 5
  8. 85.7 Claude Fable 5.1
  9. 84.7 Claude Sonnet 5
  10. 82.6 Claude Haiku 5.5
  11. 81.7 Claude Opus 4.8
  12. 79.5 Claude Opus 5.5
  13. 79.5 Claude Sonnet 5.5
  1. 1 Claude Sonnet 4.6 99.4%
  2. 2 Claude Mythos Preview 93.8%
  3. 3 Claude Opus 5 93.8%
  4. 4 Claude Opus 4.7 89.3%
  5. 5 Claude Opus 4.5 88.4%
  6. 6 Claude Opus 4.6 88.3%
  7. 7 Claude Fable 5 85.7%
  8. 8 Claude Fable 5.1 85.7%
  9. 9 Claude Sonnet 5 84.7%
  10. 10 Claude Haiku 5.5 82.6%
  11. 11 Claude Opus 4.8 81.7%
  12. 12 Claude Opus 5.5 79.5%
  13. 13 Claude Sonnet 5.5 79.5%
#ModelScoreSettingDateSource
1 Claude Sonnet 4.6 99.4% Feb 2026 Model card
2 Claude Mythos Preview 93.8% Apr 2026 Model card
3 Claude Opus 5 93.8% Jul 2026 Model card
4 Claude Opus 4.7 89.3% Apr 2026 Model card
5 Claude Opus 4.5 88.4% Nov 2025 Model card
6 Claude Opus 4.6 88.3% Feb 2026 Model card
7 Claude Fable 5 85.7% Jun 2026 Model card
8 Claude Fable 5.1 85.7% Sep 2026 Model card
9 Claude Sonnet 5 84.7% Jun 2026 Model card
10 Claude Haiku 5.5 82.6% Oct 2026 Model card
11 Claude Opus 4.8 81.7% May 2026 Model card
12 Claude Opus 5.5 79.5% Sep 2026 Model card
13 Claude Sonnet 5.5 79.5% Sep 2026 Model card

2 models have held number one since its release in Nov 2025

Lab claims count from the day the card came out, independent runs from the model’s release.

Anthropic
Lab claims Passed 85% in Nov 2025, by Claude Opus 4.5
Claude Sonnet 4.6 99.4%

The best score passed 85% in Nov 2025

All the scores we have, one dot each, so a model tested at several settings appears more than once.

Reported by the lab in its model card Best so far
0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Sonnet 4.6 Claude Opus 4.5 0% 20% 40% 60% 80% 100% 2026 Near the ceiling (85%) Claude Sonnet 4.6 Claude Opus 4.5

Labs report it only in system cards

We read the headline results of 479 launch posts from 51 labs, and none of them list this benchmark. Some system cards do, below.

17 system cards report it, from 1 lab

Besides launch posts, we read the system cards, model cards and technical reports labs publish with a release. They report far more, including safety and frontier-risk tests.

DateDocumentScoreAs the document writes itSection
7 Oct 2026 Claude Haiku 5.5 system card 82.6 Malicious computer use, refusal rate without mitigations, average of thinking on and off (Table 5.1.2.A) Safety
28 Sep 2026 Claude Sonnet 5.5 system card 79.5 Malicious computer use, refusal rate without mitigations, average of thinking and no thinking (Table 5.1.2.A) Safety
22 Sep 2026 Claude Opus 5.5 system card 79.5 Malicious computer use, refusal rate on 112 tasks without mitigations, thinking on (Table 5.1.2.A) Safety
1 Sep 2026 Claude Fable 5.1 system card 85.7 Malicious computer use, refusal rate on 112 tasks without mitigations, thinking on, row labeled Mythos 5.1 (Table 5.1.2.A) Safety
24 Jul 2026 Claude Opus 5 system card 93.8 Malicious computer use, refusal rate on 112 tasks without mitigations, average of thinking on and off (Table 5.1.2.A) Safety
30 Jun 2026 Claude Sonnet 5 system card 84.7 Malicious computer use, refusal rate on 112 tasks without mitigations, average of thinking and no thinking (Table 5.1.2.A) Safety
9 Jun 2026 Claude Fable 5 system card 85.7 Malicious computer use, refusal rate on 112 tasks without mitigations, thinking only, row labeled Mythos 5 (Fable 5 inherits these behaviors per section 5) (Table 5.1.2.A) Safety
28 May 2026 Claude Opus 4.8 system card 81.7 Malicious computer use, refusal rate on 112 tasks without mitigations, average of thinking on and off (Table 5.1.2.A) Safety
16 Apr 2026 Claude Opus 4.7 system card 89.3 Malicious computer use, refusal rate on 112 tasks run with thinking on and off, without mitigations (Table 5.1.2.A) Safety
7 Apr 2026 Claude Mythos Preview system card 93.8 Malicious computer use, refusal rate, without mitigations (Table 8.3.1.2.A) Safety
17 Feb 2026 Claude Sonnet 4.6 system card 99.4 Malicious computer use, refusal rate on 112 tasks across extended and standard thinking, no mitigations (Table 5.1.3.A) Safety
5 Feb 2026 Claude Opus 4.6 system card 88.3 Malicious computer use, refusal rate on 112 tasks across extended and standard thinking, without mitigations (Table 5.1.3.A) Safety
24 Nov 2025 Claude Opus 4.5 system card 88.4 Malicious computer use, refusal rate on 112 tasks, without mitigations (Table 5.1.3.A) Safety
5 Aug 2025 Claude Opus 4.1 system card addendum Malicious applications of computer use, no number printed. Text says compliance similar to Claude Opus 4 (3.1) Safety
22 May 2025 Claude Sonnet 4 system card Malicious applications of computer use, no number printed, qualitative findings only (3.1) Safety
22 May 2025 Claude Opus 4 system card Malicious applications of computer use, no number printed, qualitative findings only (3.1) Safety
24 Feb 2025 Claude 3.7 Sonnet system card Computer use malicious use evaluation, no number printed, qualitative findings only (4.1) Safety

More in Safety and alignment

All 444 in Safety and alignment
  • FORTRESS In use Paired adversarial and benign national-security misuse prompts. Rubric-graded risk score on adversarial and over-refusal score on benign prompts. Claude Opus 5.5 5.78%
  • StrongREJECT In use Forbidden-prompt jailbreak benchmark with a rubric grader that scores how useful a non-refusal actually is. Inkling 98.6%
  • AgentDojo In use Agents do tool-use tasks while tools and data carry prompt injections. Attack success rate counts hijacked runs. Grok 4 Fast 0%
  • MASK In use Measures whether a model contradicts its own elicited beliefs under pressure to lie. Honesty separated from accuracy. Muse Spark Open 98.1%
  • Gray Swan IPI Arena benchmark In use Indirect prompt-injection attacks from public red-teaming competitions replayed on tool, coding and browser agents. Attack success rate at k attempts. GPT-6 Astra 8.50%
  • Agentic misalignment suite (Anthropic) In use Agents in simulated company settings face threats or goal conflicts that invite blackmail, leaking or sabotage. Score is the misaligned action rate. Claude Sonnet 4.5 0%